Training Data

Aaron Levie on Where AI's Value Actually Lands

Aaron Levie· Cofounder and CEO at Box
·~65 min·English·Sequoia Capital
AgentsAI CompanyLLMInferenceBusiness Strategy
TL;DR

Box CEO Aaron Levie argues the durable value in enterprise AI is not the model but the applied layer that bridges frontier intelligence to real workflows, and he bets most enterprise tokens will soon run background tasks users never started.

01Core Mental Model

The Bridge

A model can be superintelligent and still be unable to touch a bank's legacy systems, permissions, and human handoffs on its own, so the real product is the bridge that carries intelligence into the workflow.

the models will be insanely valuable but the application of bringing those models into real workflows in banking and life sciences and healthcare and government that's just going to be a lot of software

— Aaron Levie, Training Data
Key Insight
The labs' incentive is to build the model, not to grind through the five or ten operational blockers sitting inside each customer's workflow. That grind, unglamorous and per-customer, is exactly what leaves the applied layer defensible.
02The Precedent

Snowflake Was Not Obvious

Infrastructure creating trillions does not crowd out the software built on top of it; AWS did not preempt Snowflake, and frontier models will not preempt the applied layer.

there's also trillions of dollars of value in software that only exist because of that infrastructure

— Aaron Levie, Training Data
Key Insight
The pattern is not luck. A general-purpose platform always leaves domain-shaped gaps that only vertical software can fill, so betting the platform will eat everything ignores how the last five decades of software actually played out.
03What Box Built

The Harness

A tuned agent that knows your file system, permissions, and search retrieves, reranks, and reads like an expert user, beating the same question handed to a raw model API.

effectively it's a harness for asking questions of a large data set

— Aaron Levie, Training Data
Key Insight
Box will even hand the labs its domain data, because the moat is not secrecy but being eval-maxed to one workflow. A single general model cannot be tuned to every enterprise's private heuristics at once, which is why the wrapper keeps its edge.
04Market Structure

No Lab Takes 95%

When many models are close and no single lab captures the value, the customer wants a broker that is indifferent to which model runs a task and just optimizes cost at a fixed accuracy.

the only thing I probably wouldn't bet on is just okay, one or two labs get 95% of the value creation. I think there's just going to be a much more dynamic environment

— Aaron Levie, Training Data
Key Insight
A model seller cannot credibly be that broker, because it has a preference. And the token subsidies that hide a model's true cost are temporary, so as prices fall toward commodity, the neutral operator's structural advantage only grows.
05Cost Dynamics

Both Curves Go Up

Closed frontier revenue and open-weight usage can both grow at once, because a frontier model orchestrates while matured, stable use cases get peeled off to a cheaper or open model.

you might have blended 50% spend on each but 10 times the amount of tokens, you know, on the open weights model

— Aaron Levie, Training Data
Key Insight
This dissolves the closed-versus-open, one-must-win framing that confuses observers. The two are complementary layers of a single pipeline, so exponential growth in one is not evidence against the other.
06Diffusion

Why Code Went First

Coding diffused fastest because its entire value is text a model can generate, judged by the most technical users alive, while most knowledge work sits farther down that likeness curve.

the utility of code is almost 100% represented by the amount of text that you can generate

— Aaron Levie, Training Data
Key Insight
The lesson for founders is to pick work whose output is mostly text and whose users can self-serve. Everything else needs change management, access plumbing, and patience, which is precisely the applied-layer grind Levie keeps returning to.
07The UI Shift

Tokens You Never Kicked Off

The dominant enterprise interface stops being a chat box you prompt and becomes a queue of results to review, as most tokens run background tasks no user ever started.

in five years from now I would bet like 90% of all tokens in the enterprise are things that a user never kicked off and they just see a result

— Aaron Levie, Training Data
Key Insight
If most tokens run in the background, the winning interface belongs to whoever understands the specific process, because a queue or dashboard has to name the right buttons for that workflow. The horizontal chat surface cannot.
08The Bet

Whoever Reaches the Customer Wins

When AI makes building cheap, distribution becomes the scarce advantage, and getting to the lawyer, the bank, and the sales team takes longer than Silicon Valley expects.

in a world where AI builds things so much faster. Then probably the shift moves to whoever can actually get it to the customer is in the best position

— Aaron Levie, Training Data
Key Insight
This moves the moat off model quality. Since capable models will be everywhere, the durable edge becomes the unglamorous work of diffusion: domain expertise, patience, and pushing into regulated industries one customer at a time.